GIRA

Inspiration Type 2 diabetes treatment is still largely one-size-fits-all, yet two patients with identical HbA1c values can have completely opposite responses to the same drug based on their genetics. Patients already carry CGMs, wear WHOOP bands, and have 23andMe files sitting unused — we wanted to build the tool that actually connects those streams to clinical decision-making. What it does GIRA takes a patient's 23andMe raw file, CGM history, WHOOP wearable data, and intake form and produces a structured clinical brief with medication recommendations, PGx safety flags, and live citations from ClinVar and PubMed. Deterministic safety gates — statin myopathy risk, metformin reduced efficacy, clopidogrel poor metabolizer status — always fire before any LLM output, so the model can never talk its way past a hard clinical flag. How we built it We built Python parsers for each data stream, a FastAPI backend running a tool-calling agent loop on NVIDIA NIM Nemotron, and a Next.js provider UI with tabs for genome upload, intake, CGM metrics, WHOOP metrics, and the final brief. The agent calls discrete tools — genome lookup, wearable loading, live literature search — and synthesizes across all of them into a single clinician-ready output. Challenges we ran into The hardest problem was data heterogeneity: real 23andMe exports, CGM JSON, and WHOOP responses all arrive in different shapes, and the coercion logic to handle that without silently dropping data took more iteration than anything else in the build. Keeping the LLM grounded in a patient's actual genotype — rather than generating plausible-sounding but wrong pharmacogenomic reasoning — required careful prompt engineering and the deterministic safety layer upstream of every generation step. Accomplishments that we're proud of We're proudest of the safety architecture — the decision to make certain clinical flags deterministic and model-proof was the most important call we made, and it's what makes the system meaningfully trustworthy rather than just impressive. We're also proud that every recommendation has a live citation trail, and that four completely independent data streams actually come together into one coherent brief. What we learned Integrating multi-modal health data is less a modeling problem and more a schema negotiation problem — most of the hard work lived in the parsers, not the model. We also learned that in clinical AI, what a system refuses to do matters as much as what it does. What's next for GIRA The immediate next steps are live LibreView and WHOOP OAuth so data flows in real-time, and longitudinal tracking so clinicians can see whether recommendations are actually moving HbA1c and time-in-range over time. Longer term, we want FHIR-compatible export for Epic and Athena integration, and a prospective validation study with a diabetes clinic to compare GIRA-guided care against standard-of-care outcomes.

Built With

• 23andme
• and-nvidia-nim-nemotron
• clinicaltrials.gov
• export
• fastapi
• next.js
• pubmed
• python
• raw
• rxnorm
• sqlite
• typescript
• with-live-integrations-against-clinvar